The Agent Skills Directory

[COMMAND_EXECUTION]: The documentation explicitly describes how to utilize radare2's command parser to execute arbitrary shell commands. Specifically, it notes that any command containing a pipe (|) or an exclamation mark (!) bypasses normal command handling and executes directly in the system shell.
[PRIVILEGE_ESCALATION]: The skill instructs users to set http.sandbox=false to enable features such as project saving (Ps). This configuration removes the primary security boundary intended to restrict the tool's access to the host file system and system commands.
[REMOTE_CODE_EXECUTION]: The instructions detail how to start an unauthenticated HTTP server (=h) that accepts raw command input. The combination of an exposed network interface and the documented shell execution bypasses poses a high risk of remote code execution if the server is not restricted to localhost.
[INDIRECT_PROMPT_INJECTION]: The skill performs static analysis on external binary files provided at runtime (Ingestion points: radare2 <file>, curl POST ... <binary>). These files represent untrusted data that could contain malicious strings or metadata designed to influence agent behavior. The analysis engine provides high-privilege capabilities including shell execution and file writing (Capability inventory: !, |, wx, Ps). While certain SQL sub-tables use charset validation, there are no global boundary markers or sanitization layers for general commands.
[EXTERNAL_DOWNLOADS]: The skill provides commands to download the radare2 engine and the r2xsql plugin artifact from the official radareorg organization repository on GitHub.

connect